By data maturity, we mean here an organisation’s ability to provide, for a decision or a service, data that someone answers for, whose quality is known and that can be obtained within a predictable time. We propose to place it through tests: four real requests addressed to the organisation, one per dimension, whose answer, lead time and author are observed. The dimensions chosen (responsibilities, quality, platform, uses) are those of the baseline assessment described on our data expertise page. Each gap observed designates a workstream. The grid is a method we propose; it does not rely on any published maturity model.
To follow these tests, imagine a commercial vehicle rental company whose executive management is hesitating over funding a new data platform, and first wants to know where its data maturity stands.
Testing data maturity through real requests
The grid in this article rests on one procedure, the test. A test is a request that an executive or a project manager could genuinely make, addressed without preparation, about a specific use. We note who answers, how long it takes, and what the answer rests on.
This last point gives the state of the dimension. The answer may rest on a person’s memory: it varies with the person asked and leaves with them. It may be written: a document carries it, without anything checking that it is applied. It may be verified: a mechanism applies or measures it, whether a configured access right, a check that runs or a record of decisions.
The tests concern two or three chosen uses, because the state can differ from one to another. For the rental company in our example, these would be the fleet utilisation rate presented each month to management, rental invoicing and vehicle maintenance. The result is recorded use by use, with the state observed in each dimension and the fact that justifies it.
A disputed definition puts responsibilities to the test
The test consists of asking who sets the definition of an indicator, and who would decide if two divisions calculated it differently. The fleet utilisation rate lends itself to this. Branches may count the days on which a vehicle was out, finance the days invoiced.
When the answer rests on memory, several names are mentioned, or that of the report’s author. When it is written, an owner appears in a document and the definition in a glossary. It is verified when a record shows a dated decision by that owner on that definition, and the data catalogue refers to it. The roles this test looks for are compared in our article data owner and data steward.
The figure to be justified, a test of quality
The request starts from a figure in the latest management report. Where does it come from, what checks has it gone through, and who was notified the last time it was wrong?
From memory, quality is a reputation: a given file is considered dubious, and errors are reported by those who use it. In the written state, rules exist with their threshold, and a measurement was taken once, on a known date. In the verified state, checks run in the processing pipelines and alert a named person, who can show the latest anomaly and its closure.
Tracing the figure back to its source relies on data lineage, when it is maintained. If the test reveals recurring defects with no known cause, it gives way to a longer examination, the data quality audit.
Requesting new data to judge the platform
The request: to have, for an analysis, a source that is not yet exposed. The rental company would, for example, like to match the mileages transmitted by vehicles with the rental contracts. We observe whom the request is addressed to, which route it takes and when the data arrives.
In the first state, a person who knows the database produces an extract, to be redone the following month. In the second, a request process exists and data sets are described, but loads remain manual or fail without an alert. The verified state is recognised by an automated, monitored pipeline, a measured availability lead time and spending that can be attributed to a use. These markers overlap with the indicators on our data platform page.
Uses are observed in meetings
The last request is to attend a meeting at which figures are presented, or to read its minutes. Is the time spent reconciling figures or deciding? Does each participant come with their own spreadsheet?
If each division brings its extracts and the meeting compares them, use rests on everyone’s memory. It is written when official dashboards exist, with defined indicators, while parallel extracts continue to circulate. It is verified when usage statistics show which reports are used, the authoritative content is flagged as such and the recorded decisions cite the indicators on which they rest. The single definition of indicators and the certification of content are covered on our business intelligence and analytics page.
Choosing the workstream from the furthest upstream gap
The four dimensions depend on one another, in one direction. A quality rule needs an owner to set its threshold. An automated pipeline faithfully transports data whose meaning no one has settled. A dashboard inherits everything that precedes it. We therefore suggest dealing first with the gap located furthest upstream, even if it is not the most visible.
If responsibilities rest on memory, the first workstream is data governance, limited to the uses tested, whatever the state of the platform. If they are written down and quality remains a reputation, the work concerns rules and checks, the subject of our data quality page. When these two dimensions hold and each new piece of data requires a manual extract, investment in the platform is justified. If all this is in place and meetings remain devoted to reconciling figures, the gap lies in uses.
The verified state is not the target for all data: it is justified where a use requires it. The survey describes only the uses tested, at the time they were tested. When it prepares artificial intelligence projects, it is supplemented by the one described in our article on enterprise AI maturity.
For the rental company in our example, the first test therefore precedes the decision on the platform: as long as the definition of the utilisation rate has no owner, a faster pipeline will deliver a disputed figure faster.